Agreement Analysis on Angle Characteristics With Automated Gonioscopy

Précis: Automated gonioscopy is a recent method of recording angle pictures. Our study shows that agreement between observers is moderate in most categories used in clinical practice, underlying its clinical usefulness but also room for improvement. Introduction: Iridocorneal angle (ICA) imaging methods have been recently developed to record gonioscopic images. The purpose of this study was to perform an interrater agreement analysis of ICA photographs. Methods: Multicentric cross-sectional observational study. Consecutive patients in 2 ophthalmology departments underwent automated gonioscopy. One picture per quadrant from each eye was selected for randomization. Pictures were graded and analyzed by 4 masked glaucoma specialists regarding angle opening, width, Shaffer and Spaeth gradings and other findings. Fleiss’ κ statistics was performed to assess interrater agreement. Results: A total of 50 patients were recruited, with the sample containing a variety of diagnosis including pseudoexfoliation (22%), angle-closure suspect (12%), and pigmentary glaucoma (4%). The sample included phakic (68%) and pseudophakic patients, as well as cases with gonio-implanted surgical devices (10%). There was moderate agreement on angle opening, angle width, presence of angle vessels, and peripheral anterior synechiae (κ=0.435, 0.511, 0.558, 0.488, respectively; P<0.01). Fair agreement was observed regarding Shaffer grading, site of iris root insertion, angle pigmentation features, and the presence of iris processes. Expectedly from a 2-dimensional picture, the configuration of peripheral iris was found to have a poor agreement. Subset analysis on higher quality pictures seemed to improve agreement on pigment but did not further improve on the overall good agreement on angle opening status. Conclusions: Our study shows that automated gonioscopy provides moderate agreement on most clinically relevant features. Binary characteristics such as angle opening or PAS seem to be more robust than more complex angle classification parameters.

T he anatomy and features of the iridocorneal angle (ICA), the angle between the posterior surface of the cornea and the anterior surface of the iris, are paramount in the evaluation of a glaucoma patient. 1 Gonioscopy is the method by which the ICA is evaluated, and it is a mandatory step in glaucoma consultation. It is usually performed using special types of contact lenses with mirrored facets which allow the clinician to observe the ICA structures (indirect gonioscopy) and its features, namely Schwalbe line, the trabecular meshwork, the scleral spur, ciliary body band, and iris root. Manual gonioscopy allows for detailed visualization of the ICA and its characteristics, such as width, pigmentation, iris base conformation, the presence of pathologic changes such as neovascularization or peripheral anterior synechiae (PAS), angle recession, which are essential in determining a precise diagnosis and tailoring therapy to each patient, in addition to further clinical variables such as age, staging of the disease, among others. However, this technique may be uncomfortable for the patient, requires experience, has a steep learning curve, and it is sometimes troublesome to perform detailed and minute observations of specific ICA features, especially if patient cooperation is not ideal. This leads to the omission of this essential step by some clinicians, which can cause a diagnosis to be overlooked or delayed. 2 Moreover, barring the use of slit-lamp-mounted cameras, which do not usually have enough resolution to adequately photograph gonioscopic observations and require manual focus, manual gonioscopy seldom allows the clinician to document these observations, perform a post hoc analysis and compare images to assess changes over time.
To address these disadvantages, other angle imaging methods have been developed such as ultrasound biomicroscopy (UBM) and anterior segment optic coherence tomography (AS-OCT) which will be discussed later. The EyeCam equipment (Clarity Medical Systems, CA) is a camera similar to the RetCam used for retinal imaging, which is able to take wide photographs of the angle structures and has shown good agreement with manual gonioscopy and standard goniophotography. [3][4][5] Automated or semiautomated gonioscopy with the GS-1 Gonioscope (NIDEK Co., Gamagori, Japan) is a more recent development that uses a 16-facet optical contact prism which takes photographs of 16 contiguous sectors of the ICA allowing for 360-degree visualization of the angle in < 1 minute. 6 This DOI: 10.1097/IJG.0000000000001940 equipment is capable of automatically adjusting focus on the angle structures 7 and takes simultaneous pictures per quadrant with different focal depths. They are immediately available to the clinician. Moreover, the software can reconstruct linear or circular stitching images which gives the ophthalmologist an approximate view of the real 360-degree configuration of the ICA. Automated gonioscopy has already shown promising results regarding an agreement with manual gonioscopy on angle opening and grading, 8 pigmentation of the trabecular meshwork, 9 and the identification and assessment of angle surgical devices. 10 This advancement allows for post hoc analysis of each picture and longitudinal assessment of angle changes, as it is possible with other ancillary tests used in glaucoma. Despite these advancements in angle imaging, no system has proven clinical validity. The purpose of this study was to assess the interrater agreement of angle pictures regarding ICA features in a real-life clinical setting.

METHODS
Consecutive adult patients from the Department of Ophthalmology, Hospital de Santa Maria, CHULN, Portugal, and Eye Clinic Genoa, Genoa, Italy, consented and voluntarily participated in this study. Local ethics committee approval was obtained, and this study adhered to the tenets of the Declaration of Helsinki. Inclusion criteria included patients over 18 years of age willing to participate. Patients unable to cooperate were not included, as well as those with changes in media transparency, such as corneal opacities. Patients underwent automated gonioscopy using the GS-1 Gonioscope (NIDEK Co.), which was performed by trained technicians at each site.

Image Acquisition
The image acquisition protocol and equipment details have been described elsewhere. 6 In summary, topical anesthetic drops were administered to the patient's eyes, the patient was adequately positioned in front of the machine's prism, which was coupled with a lubricating gel (GenTeal Gel; Alcon Laboratories, Fort Worth, TX). The prism is then positioned in contact with the cornea, at which point the device indicates whether the prism should be further advanced or retracted to achieve the best focus on angle structures. At the moment this best focus is achieved, the GS-1 Gonioscope automatically starts image acquisition in a circumferential fashion, taking multiple pictures at different focus depths per each 1 of the 16 sectors. Each picture is 4.6×3.4 mm in area, comprising an ∼ 30-degree view of the angle since there is some degree of overlap between adjacent images.

Image Selection
An "unmasked" investigator, who was not an evaluator, manually selected a single best-focused image from each of the 4 main angle areas (temporal-inferior, nasal-inferior, nasal-superior, and temporal-superior; Fig. 1) from a single eye per patient. Each of the 4 images selected per patient was randomized in an anonymized database. Quality screening for image quality was performed before image analysis, as explained elsewhere, 8 to ensure that only images adequately focused on the ICA were selected, in which any ICA structure was visible, regardless of angle status. Pictures showing angle closure were still included as long as the good focus was attained at the iridocorneal interface. Images with no discernible angle structures, either due to significant blurring, inadequate focus, or artifacts, were not included.  Table 1, such as image quality, angle opening, Shaffer grading, 11 Spaeth grading, 12 Scheie grading of trabecular pigmentation, 13 the presence of PAS, angle vessels, iris processes, among others.

Statistical Analysis
Continuous variables are expressed as mean ± SD. Interrater agreement analysis on discrete variables was performed using Fleiss' κ statistics. The agreement was expressed as κ-value and 95% confidence interval. Agreement level was categorized according to the Landis-Koch score 14 (Table 2). SPSS (version 26.0; IBM Corp., Armonk, NY) was used to perform statistical analysis. A P-value of 0.05 was used for statistical significance.

Sample Profile
A total of 50 patients were included in this study. Detailed patient demographics and clinical data is presented in Table 3. Most patients were Caucasian (n = 49; 96%) with dark colored iris (n = 46; 92%), with a mean age of FIGURE 1. Circular stitching of an automated gonioscopy examination of the patient included in this study, in which a large superotemporal surgical iridectomy is visible, adjacent to a trabeculectomy-related sclerostomy. The best-focused image from each of the iridocorneal angle areas between the cardinal sectors (S, superior; N, nasal; I, inferior; and T, temporal) for each patient were selected for analysis by the raters. Figure 1 can be viewed in color online at www.glaucomajournal.com.

Agreement Analysis
A total of 200 pictures were therefore analyzed and graded by the 4 raters. Detailed Fleiss' κ statistics on all variables are available in Table 4. The first variable assessed was image quality, using a Likert-scale grading style: grade 0 (bad) for blurred images in which structures are difficult to discern; grade 1 (medium) for images with only slight blurring but discernible features; grade 2 (good) for clearly focused images in which structures were easily identified. Agreement for image quality was poor but statistically not significant (κ = 0.022; P = 0.32). The mean gradings for image quality were 1.71 ± 0.51 for rater A, 1.35 ± 0.53 for rater B, 1.65 ± 0.50 for rater C, and 0.67 ± 0.67 for rater D. Agreement level on most variables was fair to moderate. Both angle opening and angle width showed a moderate level of agreement (κ = 0.435 and 0.511, respectively; both P < 0.01). Agreement on Shaffer grading was globally fair (κ = 0.268; P < 0.01), with the best results on either end of the grading scale (closed, κ = 0.363; wide open, κ = 0.438; both P < 0.01). A similar grading profile was observed with Spaeth grading regarding the site of iris root insertion (globally, κ = 0.218; above Schwalbe line, κ = 0.617 and on the ciliary body band, κ = 0.351; all P < 0.01). Agreement on the configuration of the peripheral iris, the presence of plateau iris, and pigmentation over Schwalbe line was poor (κ < 0.20), although these results were statistically not significant for the last 2 variables (P > 0.05). Agreement for the degree of trabecular pigmentation according to Scheie system was globally fair (κ = 0.396; P < 0.01), but a gradient of increasing agreement with the increase of the degree of pigmentation was seen, with a moderate level of agreement for moderate pigmentation and good agreement for intense pigmentation (κ = 0.445 and 0.719; both P < 0.01). Angle findings such as vessels and PAS scored moderate-level κ's (κ = 0.558 and 0.488, respectively; both P < 0.01), while agreement on the presence of iris processes was fair (κ = 0.354; P < 0.01). Furthermore, all raters correctly identified all surgical devices present (XEN gel stents visible; n = 2) and posttrabeculectomy status (n = 2) (Figs. 3, 4). Since the number of pictures in which these surgeries were visible was so small, Fleiss' κ statistics was not performed.

Subset Analysis
To determine if the perception of image quality influenced agreement on the remaining variables, 2 further subsets of images were created. Subset 1 was composed of 86 images which were rated as "good" by at least 2 observers and had no "bad" ratings, while subset 2 was composed of 49 images rated as "good" by at least 3 raters and had no "bad" ratings. Detailed results are visible on Table 5. Agreement on angle opening was similar in all image sets (κ = 0.435, 0.483, and 0.414 for whole agreement, subsets 1 and 2, respectively; P < 0.01 for all) but decreased regarding angle width (κ = 0.511, 0.531, and 0.330; P < 0.01 for all). Agreement on pigment over Schwalbe line, pigmentation normality, Scheie grading of trabecular pigmentation, and on the presence of angle vessels all increased. There was no significant change in agreement regarding Shaffer grading, Spaeth grading parameters (except for angle width), and PAS.

DISCUSSION
The purpose of this study was to assess interobserver agreement regarding several ICA features using a method of automated gonioscopy with the GS-1 Gonioscope. This equipment has previously shown promising results in identifying angle variables such as trabecular pigmentation and characterizing surgical devices and has the benefit of rapidly recording 360-degree angle images for post hoc analysis. 6,9,10 Observers in this study were asked to rate ICA pictures regarding several distinct angle variables.
Our study showed fair to moderate agreement levels on most ICA features analyzed. Agreement was better in binary variables, such as angle opening (open vs. closed; few images were rated as "not sure"), angle width (narrow vs. wide), angle vessels, and PAS (not present vs. present), whereas in variables which require more minute and detailed analysis angle features or hold multiple categories, such as Shaffer grading, Scheie trabecular pigmentation grading, and pigmentation over Schwalbe line, agreement was worse. Nevertheless, we also observed that as some features became more extreme (ie, Shaffer opening grading 0 or 4, Scheie pigment grading 3 or 4), agreement increased because these angle changes were more apparent and harder to miss.
In clinical practice, the most important gonioscopic variables are those which directly impact patient management. Angle opening, and the presence or absence of  iridotrabecular contact, is possibly the most relevant variable, not only in the nosology of glaucoma, but also because it dictates whether a particular patient may benefit from distinct interventions, such as cataract surgery, a peripheral iridotomy, or angle surgery. A previous study by Teixeira et al 8 using an earlier prototype version of the GS-1 Gonioscope showed that automated gonioscopy underperformed in identifying angleclosure versus manual gonioscopy. It is unclear if the newer commercially available equipment we used would perform any differently regarding this variable, as we did not perform a comparison with manual gonioscopy. However, it has also been documented by Quigley et al 15 that fewer than half of glaucoma patients undergo regular manual gonioscopy, a figure which could conceivably be even lower considering the global pool of ophthalmology patients. Given that automated gonioscopy may be relatively easier to perform than manual gonioscopy and requires a less steep learning curve, an argument could be made that this technique could help identify angle-closure cases which could otherwise be missed. The presence (and extent) of PAS, angle vessels, and intense trabecular pigmentation can also direct the clinician towards different diagnoses and therapeutic approaches, such as using laser trabeculoplasty in intensely pigmented angles. Our results suggest that automated gonioscopy performs fairly well in documenting these variables and should be able to demonstrate, for example, angle width changes after a surgical or laser intervention, changes in PAS formation in an uveitic patient over time, or abnormal angle vessels in neovascular cases. Moreover, all surgical devices and posttrabeculectomy signs present were correctly identified by all raters. This implies that this equipment may be useful in recording and analyzing angle devices (such as minimally invasive glaucoma surgery) or iatrogenic damage to the angle over time, as has been suggested previously. 6,10 To our knowledge, the publication by Teixeira et al 8 is the only study to have featured an analysis on interrater  16 These differences may be difficult to interpret considering differences in methodology, such as the monocentric nature of both studies. Even though no images without discernible ICA features were included during image selection, we still found a lack of agreement between raters concerning image quality. Consequently, we decided to create 2 subsets of images according to rater responses on image quality, with subsets 1 and 2 comprising angle pictures the raters agreed were of higher quality, as previously detailed. Agreement analysis on these 2 subsets showed variable changes in κ values, with no discernible overall tendency, but some variables which naturally heavily rely on image quality, such as pigmentation over Schwalbe line, Scheie grading of trabecular pigmentation, and presence of angle vessels showed increased in agreement. These results may imply that, despite an absence of agreement on image quality among the raters, the perception of image quality was not a dominant factor in influencing the agreement level regarding more "macroscopic" ICA features such as angle opening, angle width, and PAS, but can influence the correct classification of more complex and detailed angle characteristics. Another possible explanation for the lack of an overall tendency in agreement change comparing the original picture set with the subsets created may be that by reducing the number of pictures in these subsets, particularly in subset 2, the statistical power of the agreement analysis may have also been decreased.
Other angle imaging methods, such as AS-OCT or UBM, can provide accurate morphologic and morphometric data on angle structures, which may be important, for example, to correctly diagnose a plateau iris, 17,18 an ICA feature our study showed to have a poor agreement (although not statistically significant) using automated gonioscopy. However, automated gonioscopy pictures hold the advantage of being a representation of what the ophthalmologist actually sees during gonioscopy, and as our study shows, it can identify and grade changes such as PAS, angle vessels, trabecular pigmentation, and surgical iatrogenesis, which neither AS-OCT nor UBM would be able to document. Compared with EyeCam imaging, which provides wider pictures of the ICA and shows good agreement regarding angle closure, 4,5 automated gonioscopy provides closer images of the angle, which may be an advantage in identifying smaller details and angle structures. One possible drawback of automated gonioscopy versus manual gonioscopy concerns the fact that it is still unknown if the apposition of equipment's prism on the corneal surface produces some indentation. However, considering that the diameter of the prism surface is similar to a 3-mirror Goldmann type lens, indentation would imply a misalignment between camera and cornea. Moreover, since there is no movement of the prism relative to the cornea during the examination, there is also no variation in the force exerted by the prism on the cornea. Accordingly, one would anticipate that a 360-degree good quality picture would imply a proper alignment and centration and thus a minimal indentation (if any). Indeed, some angle parameters, such as Shaffer and Spaeth, may be optimized by a combination of static, dynamic, and indentation gonioscopy. Assuming automated gonioscopy with the GS-1 Gonioscope is predominantly a static method of gonioscopy, this may account for only fair to moderate agreement between specialists regarding some of these angle variables.
Limitations to our study include a relatively small and homogeneous population regarding iris color and ethnicity. Also, agreement analysis in glaucoma is often difficult to assess due to the subjective nature of some ocular characteristic valued by clinicians, such as the estimation of width, the shape of the peripheral iris, the intensity of pigmentation across different ICA structures, among others. In fact, previous studies on the agreement between clinicians on optic disc photographs yielded variable results. [19][20][21] Fleiss' κ statistics depends heavily on the number of categories per variable and on the number of raters. Generally speaking, as the number of subjects in analysis increases, the more statistical power the analysis holds, and as the number of categories and raters increase, the expected agreement decreases. 22 This meant that we expected lower agreement levels on some variables than what similarly designed studies which also dealt with angle